Merging error analysis of name disambiguation based on author similarity

Zheng Xie, Jianping Li, Miao Li, Dongyun Yi, Yanqin Feng, Zonglin XiearXiv 2017 · arXiv

Abstract

In network science, constructing accurate coauthorship networks is often hindered by identity errors. While "splitting errors" fragment a single person into multiple nodes, "merging errors"—where distinct individuals are treated as one—are particularly damaging, as they distort network topology and inflate author metrics. This paper introduces a Bayesian framework to quantify these merging errors by analyzing author features like coauthor similarity. By iteratively updating the probability of identity matches and error rates, the model provides a more reliable estimate of the true number of individuals in large-scale social networks.

The Identity Crisis in Network Data

Building coauthorship networks involves mapping author names to unique nodes. However, researchers frequently encounter two primary identity errors:

  1. Splitting errors: A single individual is represented by multiple nodes (e.g., "J. Doe" and "John Doe").
  2. Merging errors: Multiple distinct individuals are collapsed into a single node (e.g., two different "John Smiths").

Merging errors are especially problematic because they artificially reduce the number of nodes, shorten average path lengths, and inflate productivity metrics. Since ground-truth data is rarely available for massive datasets, the extent of this "compromise" often remains unknown. This research proposes a mathematical framework to measure and correct for these errors.

Methodology: The Bayesian Framework

The authors propose an iterative Bayesian model that leverages observed author features—such as similarity in coauthor lists, affiliations, or email addresses—to refine identity estimates.

The process begins with a similarity matrix and uses Bayesian inference to calculate the probability that two author instances are actually the same person ($A_{ij}$), the false positive and negative rates ($e^+$ and $e^-$), and the overall probability of identity within an entity ($\rho$).

The Bayesian model workflow: similarity scores are processed to output identity probabilities and error rates.

This model simplifies previous approaches by removing complex hyperparameters, making it more practical for empirical research.

Empirical Evaluation

The model was demonstrated on three major scientific publication datasets: SCAD-zbMATH (a manually disambiguated "gold standard"), PNAS, and PRE. These datasets were processed into different versions based on how names were recorded to test the model's sensitivity to varying levels of ambiguity.

Statistical indexes of the empirical datasets used for evaluation, including zbMATH, PNAS, and PRE.

Key Insights and Results

The Information Metric ($I$)

The researchers defined an information metric, $I = 1 - (e^+ + e^-)$, to quantify the "carried information" of any given feature. If $I$ is near zero, the feature (like coauthor names) provides no useful information for disambiguation. The model revealed that using coauthor names often carries "null information," failing to reduce data uncertainty.

Hyperdegree Correlation and Data Inflation

A significant finding is that merging errors are not uniform. Authors with a high "hyperdegree"—those appearing on many papers—are much more likely to be victims of merging errors.

The model's ability to estimate the "true" number of authors ($\hat{N}$) revealed staggering levels of data inflation. In some cases, the estimated number of real individuals was over 10 times higher than the number of unique entities in the raw data. For the zbMATH dataset, the ratio of estimated authors to raw entities reached as high as 1050%, indicating a massive "hidden" population of authors.

Limitations and Future Work

While powerful, the model currently focuses exclusively on merging errors and does not address splitting errors. Furthermore, its performance is dependent on the quality of the input similarity features. Future research could extend this framework to assess the confidence levels of existing network science conclusions, determining if established phenomena like "six degrees of separation" are artifacts of noisy data.

Summary updated

BibTeX
@misc{xie2017merging,
  author = {Zheng Xie and Jianping Li and Miao Li and Dongyun Yi and Yanqin Feng and Zonglin Xie},
  title = {{Merging error analysis of name disambiguation based on author similarity}},
  year = {2017},
  eprint = {1711.01406},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1711.01406},
}
Figure